Ive been a wordpress developer for the past 20 years and then became the lead search engine optimization manager for an agency, I see a lot of things in AI that are replacing the need for me, I’d like to learn a new skill set in AI that makes me irreplaceable. Maybe I can run an ai automation agency
each problem you will need to adjust your dataset, this is why data engineers and data scientists work together. If i were a beginner i would start reading public IA notebooks in kaggle, in order to be able to create my own dataset.
why are some here trashing david ondrej? he is imparting knowledge in an easy to understand way for peeps that do not know. i wrote my first neural net from scratch in 1993 and i have been an ML practitioner since then. i can tell u that info back then was hard to come by. be grateful that u have easy access to it. if u dont like it better to move along rather than disperse caustic.
Thanks for the video, David. Really appreciated seeing such a detailed E2E example as I evaluate options for training a custom LLM. Keep up the great work!
I have seen a lot of videos on fine-tuning and read a lot, and I have to say this is one of the most lucid, explanations. By making it very concrete and showing the code and, importantly, the training data you make very clear what is going on in fine tuning great job!
He talks about this in his example where he made an app similar to Blinkist and uploaded a free ebook as his example and he mentioned a section where he added a pdf script somewhere in his Python code so that it would filter certain things in the pdf like the useless cover pages and page numbers etc etc
@@DavidOndrej What I'm looking for is an unregistered Dalle model The reason for it is that Dalle can convert simple text into extended prompts, unlike many other engines like sdxl
3:37 He is 100% correct that already fine-tuned LLMs like GPT, Claude, and even Gemini 1.5 Pro with 1m+ context, are freaking awful at trying to emulate writing styles. Worst part about ChatGPT for this purpose is that no matter how much you tell it not to, it's filled with clauses like "On the other hand,", "Finally, ", or "As a consequence" and I'll explain to it again all the reasons those phrases don't belong in a rap song.
Hey I could be wrong but I believe you'll have better luck fine tuning with a less quantized version of the model. At least 8 or 16bit would be preferable to 4. I'm not an expert on quantized models, but you lose a lot of resolution when you quantize that much and that likely makes it more challenging for the LoRA to train. Definitely correct me if I'm wrong, folks, but I think this is the case.
Yes, this is consistent with the programming adage 'optimize last'. The trade off is speed for accuracy, but refined-model accuracy will be more important in the longer term than the speed of the refinement process itself.
I love how enthusiastic you seem. You seem super interested in this stuff. AI is not my specialty, im utilizing it more as a tool for the projects I work on (security research and other stuff). Cool to see someone just being themselves :)
Nice video, down to the point and good overview. And cool your honest about how much you know. I like that. Would be interesting to have a video about creating dataset and using Agents for training.
This is a great video David, you've got yourself a new subscriber. I've been looking for some guidance on this for a while. Don't sweat on not being a complete expert on the topic; you don't need to be 100% across every aspect of a topic to point someone in the right direction. People can fill in the gaps as required.
Awesome video, glad I found your channel. Shout out from South Africa. A video on how to prepare datasets for training, like finetuning an 8b model for a MySQL database. Peace brotherman.
vicious circle. give llm a little data and say simulate. llm uses trained data to simulate. user takes simulated data and does the same in another llm. very little new data has been added to the system.
Hey Ondrej! I think this might be a stretch of the topic, but is it possible to use an llm like llama 3 and fine tune it to respond in another language or would it be necessary to train an llm from scratch for this?
This is a very useful topic, in the future we can train our datasets to specifically use them for different applications, particularly in healthcare or other institutions, benefiting people. Hopefully, a next topic will be about how to create your own datasets. Thanks for the explanation
Actually GPT-4 can be fine-tuned by the user, it's done within the openai API and of course used by it's API later on. It obviously has downsides, like the model is still invoked on the OpenAI servers and they are collecting all the data which goes through it (no privacy), but it is possible :)
Hey David, great video and great explanation. Please make a tutorial on how to generate dataset using LLM. For my use case, I have a classification problem and the class imbalance is severe. for the minority classes, I want to generate more meaningful samples using LLM and then build an LLM model to do text classification on the dataset. Any suggestion on achieving this would be great!
I didn't understand almost anything, but the video is excellent. You really tried to help. The problem is mine, which needs to have an entire course on "who knows what" to understand that. Perhaps beginning to learn Python, Linux, and data engineering.😅
You can definitely fine-tune ChatGPT 3.5 and you can also ask open AI to invite you to their private waitlist to be able to fine tune GPT 4. So it is definitely possible.
10:29 that would be great tbh, using agents to make dataset to finetune the model is just like inception, you can also make agents to prepare dataset for other agents to create dataset to finetune the model (inception level 2) or make agents to prepare dataset for agents to prepare dataset for finetuning the model which will be used to prepare dataset for agents to prepare dataset............
Hi David I think it would be a fantastic idea to do a video on how to make a dataset for llama 3.1 on one complex API. I repeat, complex, long API with multiple inputs and outputs. Every time I see an example the APIs are too simple but real world apis are long and complex they might not fit in a prompt
Hi, I didn't see how this is domain specific, as I tested the 2 prompts on llama3-8b-instruct itself, the ouput is correct as well. I've tried to fine-tune the instruct model with some Q&A dataset, and it can correctly answer original questions from the dataset, but cannot answer them correctly if I paraphrase the question. Would be great if you can share more on domain specific perspective.
This video is exactly what I was looking for. Thank you. Now, I wish to know which hardware configuration I will need to install and use Llama 3 models locally in my own machine. Can you help me?
Could I fine tune my private llama 3 llm but also connect it to chat GPT 4 API such that it can reach out for additional information beyond what the pre-trained model and private knowledge base May have? If so would it then be using chat GPT for API almost like some people can figure rag to look at their private data sets?
I'm finding the data-set stuff very confusing. What if I want to create a data-set that's just my writing? I want the model to emulate my writing perfectly. I don't have question/answer pairs.
can you offer any advise about importing the ggufs into ollama, mine just spit out gibberish, I presume it has something to do with the modelcard but no idea
it's a good video, thanks. But there are a lot of videos about fine tuning. It would be perfect if you would create a video on how to create own data sets for fine tuning. 👍
Is the trained model able to be used with "ollama run trained_model_name"? Do I have to download it directly and put it some where for that to work? I currently have a python program setup that uses the ollama module and runs llama3. But I would like to use a fine tuned model instead as I am trying to make a Jarvis like personal assistant!
hi i have a question, what if i want to use my dataset json file into the cell instead of huggingface alpaca json. Can you give the part of the input code
Oh wow you managed to fix the fine-tuning issue? Its been a headache for the entire open source rn, because Llama 3 trained their models differently so every fine tune would end up way worse than the original base model.
If you watch the video, you will see that I openly admit that I am not an expert when it comes to fine-tuning. In fact, making this video definitely was outside of my comfort zone.
@@DavidOndrejay respect man. In fact I think you should include more fine tuning to your videos in the future. You can’t run away from fine tuning if you want A.I to move to commercial use. Llama 3 is probably the only exception in the industry rn that has everyone stumped
when are you going to make datasets for fine-tuning, I have currently data in mysql that I need to extract and create the datasets for fine-tuning llama.
Is there any free method to fine tune an large language model locall. I have a small workstation with 128GB DDR4 memory, Nvidia RTX A1000 X2 SLI VGA, AMD Threadripper process. I tried AutoTune-Advanced and LLaMA-Factory. They both failed on me. Autotrain say I dont have enough VRAM. LLaMA-Factory say I dont have CUDA. Please help me.
perfect timing i was just thinking about having multiple llama 3 versions fine tuned for specific coding projects instead of a broad coding language base. is this just a waste of time and im better off having a general coding version instead? i was considering having a few fine tuned models to imitate a development team with crew.
Its is possible to create something local with llama to run on a raspberry pi and have it check spelling and grammar and rephrase things like Grammarly
I am looking for tutorial how to generate dataset using Agents. There is no such tutorial (or I am not able to find it). It would be great to generate chat format (conversation) dataset as a response of task. So as an input you have list of task, question and then agents generate conversation to this topic.
How to train a data set which is not in the form of instructions, input, output format, lets say I want to train the model using the data from a pdf, or any other means, how can we do that, please suggest some ideas. Thanks in advance.